CFD Investigation of Large Scale Pallet Stack Fires in Tunnels Protected by Water Mist Systems

Transcription

1 CFD Investigation of Large Scale Pallet Stack Fires in Tunnels Protected by Water Mist Systems JAVIER TRELLES* AND JACK R. MAWHINNEY Hughes Associates, Inc., Baltimore, Maryland, USA ABSTRACT: A series of full-scale fire suppression tests was conducted at the San Pedro de Anes test tunnel facility near Gı jon, Asturias, Spain in February 26. The fuel was wooden pallets or a mixed load of wood and high density polyethylene pallets. Fire protection was provided by water mist systems in different configurations. Because of facility restrictions, some scenarios of great interest, such as a free burn fire, could not be investigated. However, in order to complement the experimental results, a number of computational fluid dynamics simulations were conducted on a 14 m section of the tunnel facility. The Fire Dynamics Simulator, version 4, was used for the numerical investigation. An algorithm was developed to allow the fire to spread along the top of a series of pallet loads in such a way that the measured heat release rate was reproduced. Verification and validation studies confirmed that the model predicted the measured ventilation speeds and peak temperatures. The agreement between the simulations and the field measurements was very good prior to activation of the water mist. Back-layering was modeled well. After activation of the mist, the simulations predicted a large drop in gas temperatures, and retreat of the back-layer, but under-predicted the thermal cooling by the water mist downstream of the fire. With the suppression system, high temperatures and heat fluxes were limited to the immediate vicinity of the burning pallets. The model was then used to simulate a free burn fire in the tunnel. The simulation demonstrated the catastrophic conditions created by an unsuppressed fire in a tunnel when compared against the thermally managed conditions under suppressed conditions. KEY WORDS: tunnel, water mist, CFD validation, fire suppression, flame spread modeling. INTRODUCTION AS A RESULT of a number of multiple-death fires that have occurred in the last decade in highway tunnels in Europe [1], there is growing *Author to whom correspondence should be addressed. jtrelles@haifire.com Journal of FIRE PROTECTION ENGINEERING, Vol. 2 August /1/ $1./ DOI: / ß Society of Fire Protection Engineers 21 SAGE Publications, Los Angeles, London, New Delhi and Singapore

2 15 J. TRELLES AND J. R. MAWHINNEY pressure on public highway officials to increase the level of fire protection in road tunnels. Passive fire safety systems receiving increased attention include improved insulation to protect the tunnel lining, improved smoke and fire detection, improved traffic surveillance, and automated video detection [1]. Active fire safety systems under consideration include sprinkler systems, water mist systems, and both high-expansion and compressed-air foam systems. The current investigation examines the effectiveness of high pressure water mist systems against the very large fires that can occur in tunnels. Any of the active systems proposed raise questions regarding evaluation of the benefits of the systems relative to the range of fire severity that can occur. Transportation authorities must decide the difficult question of what constitutes an adequate, affordable level of protection, both in terms of life safety and of property protection. To address some of these questions, extensive fire testing of a high pressure water mist system against heavy goods vehicle (HGV) fires was conducted in the San Pedro de Anes fire test tunnel in Spain. Data from the fire tests was reported in [2] on the heat release rates of burning wood and plastic-wood pallet mixtures under suppressed or partially suppressed conditions. The test series raised a number of questions about the nature of the impact of the suppression system on the tunnel. In order to completely understand the performance of the systems, computational fluid dynamics (CFD) modeling was used to analyze the complex turbulent fires of this test series. The goal was to illustrate the dynamic flows and the thermal environment in specific regions within the zone of influence of the fire and in regions beyond the instrumented portion of the tunnel. It is known that standard HGVs loaded with common materials such as furniture may contain significant amounts of plastic materials. The fires that can be produced by transport vehicles carrying such materials have been shown to exceed 15 MW under unsuppressed conditions [3] depending on the quantity of plastic contained in the fuel. In this investigation, the CFD model was validated against data obtained from a series of test fires in wood pallets only (referred to as standard severity fires) and test fires involving wood pallets with 16% HDPE plastic pallets (referred to as high severity fires). The goal of the current article, which only discusses the wood pallet fires, is to quantify and demonstrate the benefits of the suppression system for these fires. The reader s attention is also directed to a companion article, Mawhinney and Trelles [4], which covers certain topics not addressed here in great detail. LITERATURE REVIEW The following is a sampling of the tunnel CFD work that has been done to date. Reviews can be found in [5 12]. A great deal of the existing literature

3 CFD Investigation of Large Scale Pallet Stack Fires 151 covers ventilation studies. For example, [13] gives advice on how to perform tunnel ventilation studies in the presence of fires and then gives results from several case studies. Data obtained from the Ofenegg Tunnel fire tests were validated [14] with the commercial CFD code, Flow3D. Flow3D was also used to model the heptane tests [15] performed in Project Eureka as well as to investigate the interaction between tunnel ventilation and fire-induced flows [16,17]. Sensitivity analyses were performed and uncertainty estimates reported in the latter studies. References [18 21] used CFD to determine ventilation rates that prevent back layering while [22] compared 3D FLUENT calculations with 1D SPRINT results. An unnamed CFD model was used to look at the influence of slope on 3 MW fires within a stretch of underground roadway beneath Barcelona [23]. That study looked into the propagation and extraction of smoke (by means of semi-transversal ventilation), particularly in the initial stages previous to the activation of the smoke control system when the spread of smoke is checked only by the presence of the smoke ducts. SOLVENT [24] was developed for tunnel ventilation simulations in Phase IV of the Memorial Tunnel Fire Test Ventilation Program [25]. It is based on the general-purpose CFD code COMPACT-3D [26] and was used at the National Research Council of Canada (NRCC) in a series of tunnel ventilation studies [27 31]. The target tunnel had a short ramp-down, a long straightway, and then a short ramp up. These authors have also used fire dynamics simulator (FDS) in their series of studies of fires in tunnels. TUNFIRE was used for ventilation validation [32] using data from several tunnel studies. JASMINE-SPARTA was used to look at drag-down from sprinkler systems and to calculate water flux maps. In [33], various aspects of modeling spray barriers were investigated, including an annular barrier midway through a cylinder. This work is also significant for comparing the Eulerian and Lagrangian methods of modeling sprays. FLUENT was used to study the impact of a small water mist system on a 5 kw propane fire in a 2D tunnel 2 m tall in the absence of ventilation [34]. That study found that sufficiently high water flow rates led to conditions that would not sustain combustion. FDS and its antecedents have been used in several studies. For example, FDS was used for extensive safety studies within the tunnel networks that comprise the Gran Sasso National Laboratory in Italy [35,36]. These efforts encompass smoke transport, egress studies, and water mist protection. FDS was also used to investigate the fire that occurred in the Howard Street Tunnel in Baltimore [37,38], but Richtwasserstaat (RWS) [39] temperatures could not be achieved at structural surfaces. Ventilation issues resulting from tunnel boring operations were examined with a combination of 2D and 3D modeling [4]. In [41], FDS was used to model the gasoline fire within

4 152 J. TRELLES AND J. R. MAWHINNEY the Caldecott Tunnel, where the maximum wall temperature was 958C. FDS was used in [42] to reproduce the RWS time curve [39] in the Clyde Tunnel, where it was found that the inclination was such that a 5 MW fire had to be used to get temperatures similar to the RWS curve. In order to aid with the design of the IPS foam system, Cafaro et al. [43] used FLUENT to model complex tunnel geometries and FDS for near-field conditions. In [44], several models were used for risk-assessment studies intended to guide the decision making process in the presence of aleatory and epistemic uncertainties. A validation study [45] was performed based on data gathered from fires conducted within the YuamJiang #1 Tunnel for which FDS ver.4 results agreed very well with the data collected during that test series. The Memorial Tunnel test data were used in [46] to perform sensitivity analyses and a flame spread algorithm was developed in [47], similar to the one used in the current publication, for Runehamar fuel loads. Nmira et al. [48] represents an important advance in the modeling of fires in tunnels protected by water mist. The model they developed used Arrhenius kinetics for the heat release rate and predicted suppression/ extinction. The scale of their computational domain was relevant to transportation tunnels. Although they used simple fuels (PMMA) and only one water mist nozzle, the results were very impressive. UNCERTAINTY It is desirable to know how the uncertainty in key input variables, such as the heat release rate, is reflected in the calculated results. One way of accomplishing this is through a sensitivity analysis. Early examples with respect to tunnel simulations can be found in [49,5]. In general, the sensitivity analysis approach involves either solving extra sensitivity equations, which are not currently part of FDS, or performing extra simulations with certain variables perturbed, which greatly increases turnaround time. For these reasons it was decided to adopt the method of Hamins and McGrattan [51], which is described in the Appendix. NOZZLE AND SPRAY CHARACTERISTICS Multi-port High Pressure Nozzles Two Hi-Fog nozzles were used in the test series. As is shown in Figure 1, the 4S 1MD 6MD 1 model is a cap-protected, multi-port nozzle. Model 2N 1MD 6MD 1RE is a closed, individually thermally activated nozzle with a protective cap (Figure 2). Model 4S 1MD 6MD 1 was used in the

5 CFD Investigation of Large Scale Pallet Stack Fires 153 Figure 1. The Marioff 4S 1MD 6MD 1 HI-FOG (4S) high pressure water mist nozzle has seven orifices. Each nozzle is protected by a cap that is hydraulically released upon pressurization of the deluge zone piping. Its K-factor is 5.5 L/min/bar 1/2. Figure 2. Model 2N 1MD 6MD 1RE (2N) is a closed cap, thermally activated, multi-port nozzle with seven orifices identical to the deluge nozzle shown in Figure 1. It has a K-factor of 5.3 L/min/bar 1/2. The protective cap is hydraulically released to expose the thermal element upon pressurization of the zone piping. (The color version of this figure is available online.)

6 154 J. TRELLES AND J. R. MAWHINNEY Table 1. Summary of tunnel and system conditions for the selected tests. Test identifier Tunnel station bounds (m) System mode Target Achieved pressure pressure (bar) (bar) T act (8C) # Nozzles activated/ zone total _V T H2O (L/min) (8C) Ventilation speed a (m/s) T 1,air (8C) 1 [358 43] Hybrid / [358 43] Hybrid / [358 43] Sprinkler / c [372 4] Sprinkler b / [37 42] Hybrid / a Includes the effects of exterior wind. b The sequence indicates that the pressure dropped as additional water mist nozzles were opened by heat. c In the water mist sprinkler mode, the quantity of water flowing in the pipes is less than in the hybrid mode. As the pipes are exposed to heat from the flames, the temperature of the water inside the pipes is higher than when flow rates are higher. hybrid tests and it would be the only nozzle used in the zoned deluge mode experiments. Model 2N 1MD 6MD 1RE was used in conjunction with the open spray nozzles in the hybrid tests. It was the only nozzle used in the water-mist sprinkler mode tests. Flow data are given in Table 1. Drop Size Distribution The drop size distribution was measured by the manufacturer only for the 4S 1MD 6MD 1 nozzle [2]. It is expected that the drop size distribution for the 2N 1MD 6MD 1RE nozzle would be very similar because both nozzles have the same orifice diameters, operate at the same pressure, and have nominally the same K-factors. FDS uses a Rosin-Rammler/log-normal distribution for the drop size distribution. Refer to [52] for a complete description of the relevant formulae. Chan [53] found that this composite analytical distribution provided a good fit to traditional sprinkler data. Reference [4] explains how the distribution parameters were determined for water mist nozzles. FULL-SCALE FIRE TESTS Full-scale fire suppression tests were conducted at the San Pedro de Anes test tunnel facility near Gíjon, Asturias, Spain in February 26. The fire tests involved fuel packages representative of HGV loads. A water mist system was installed in the tunnel with two different operating modes water mist sprinkler mode, which is a zoned piping system with individually thermally activated nozzles covered by protective caps (Figure 2), and a hybrid mode

7 CFD Investigation of Large Scale Pallet Stack Fires 155 Table 2. Summary of fuel data for the fire tests that constitute the validation suite. Test identifier _Q m (MW) Fuel configuration Fuel type Fuel array dimensions (m m m) Tunnel station bounds (m) Tarp? Wind breaker? Percent left standing Euro Pallets [ ] No No Euro Pallets [ ] No No Euro Pallets [ ] No No EP & HDPE [ ] Yes No Euro Pallets [ ] No No 5 In Test 13 the fuel package was positioned 2.4 m off center. Q _ m refers to the nominal maximum heat release rate. The growth period is the time it took to get to _Q m. The spread rate is along the length of the fuel array. EP stands for Euro Pallet. HDPE stands for high density polyethylene (pallet). system, which is a zoned piping system with water mist sprinklers interspersed with open spray nozzles. The objective of the fire tests was to evaluate the effectiveness of the water mist system at controlling temperatures, reducing the severity of the fires, and preventing fire propagation in the tunnel. Of 11 tests conducted in the tunnel and fully documented, five were chosen to be simulated. Refer to Table 2 for fuel load details. The first three tests involved the same fuel package, consisting of wood europallets stacked on an elevated platform to represent a standard severity HGV load. The estimated maximum potential heat release rate for the standard severity fuel package was 75 1 MW. The pallets were constructed to an ISO standard so that dimensions, weight, and moisture content of the wood were as consistent as possible for all tests. In each of the first three tests, the wood pallets were stacked on the centerline of the tunnel directly under the middle water mist line. The pallets were uncovered and open to the ventilation air flow in the tunnel, which was in each case approximately 2 m/s. The only differences between the three tests were the details of the water mist system. In the first test, the water mist system was operated in hybrid mode at 1 bar pressure; in the second test, in hybrid mode but at 8 bar pressure. Refer to Table 1 for water mist systems data. CFD simulations were performed for all the tests shown in Tables 1 and 2. However, since only results for Test 1 are presented in this article, a discussion of the other tests is left to a subsequent publication. It was not possible to conduct a full-scale test with an unsuppressed fire in the San Pedro de Anes test tunnel, due to concerns about damaging the structure. However, a sixth simulation was performed using the HRR input measured for the severe fire load in Test , placed in the same

8 156 J. TRELLES AND J. R. MAWHINNEY position in the tunnel as the fuel package in Tests 1, 2, and 3, but with the water mist system turned off. THE TUNNEL ENVIRONMENT Good modeling practice, as recommended in [52], includes preliminary evaluations of the environment to be modeled. Table 3 summarizes these findings based on conditions up-wind of the fire location. The velocity is the ventilation rate, U 1. The Mach numbers, M 1, are suitably low for FDS simulations. Because tunnels can use jet ventilators, this is not always as given. The Reynolds numbers based on the hydraulic diameter, Re d, indicate turbulent flow, as expected. The Froude numbers based on the tunnel height, Fr 1, indicate that the tunnel ventilation can adequately control mildly buoyant flows. The back layering numbers [54]:! N BL ¼ ð 1Þg 1=3 _Q m p 1 U 3 1 H, ð1þ indicate that the extent of back layering scales as times the tunnel height (without taking the tunnel grade into account). Uncertainties were also estimated a priori. Table 4 lists them for relevant quantities as functions of the uncertainty in the heat release rate. Use will be made of these below. Table 3. Nondimensional numbers associated with conditions upstream of the fire. Run identifier U 1 (m/s) T 1 (K) _ Qm (W) M 1 Re 1 Fr 1 N BL ,,.6 923, Unsuppressed ,5,.6 922, Table 4. Uncertainties associated with _ Q m, the nominal maximum heat release rate. Test identifier _Q m x c (MW) Q _ m (m) x c (m) t c (s) t c (s) U c (m/s) U c (m/s) % to þ2% % Unsuppressed 58 15% to þ25% % T

9 CFD Investigation of Large Scale Pallet Stack Fires 157 CFD Model FDS 4..7 SIMULATIONS FDS 4..7 was used for all the simulations. FDS 4 is a 3D large eddy simulation CFD model [55,56] created specifically for studies related to fire protection engineering and fire science. It contains many sub-models and control features that allow inclusion of items such as vents and nozzles, which open and close at specified times. The following subsections discuss certain aspects in more detail. The FDS simulations were run on a cluster of Linux computers comprised of Pentium IV single processors with 4 GB of memory each and multi-core processors with access to 8 GB of memory. Each processor/core had a clock speed in the 3 GHz range. Run times ranged as long as 7 days. Modifications to FDS4 Overall, the official release of FDS 4..7 has been used for all the calculations. However, it has one inadequacy that was rectified for this study. FDS4 calculates the log-normal standard deviation, r LN, from the Rosin-Rammler exponent, g RR, by imposing slope continuity at the intersection between the two branches of the composite distribution. Unfortunately, there is no justification for this smoothness and it has been found to provide a poor prediction of the diameter at the 1% cumulative volume fraction point (Dv1) for water mist applications. Figure 3 compares Rosin-Rammler/log-normal fits with measured data that were obtained in a separate study [57,58] for a Marioff 4S 1MC 8MB 11 nozzle. The smooth log-normal branch advocated by FDS4 provides a poorer representation of the data than does the nonsmooth branch obtained by the methods presented in [4,52]. This was found to be the case with all of the water mist nozzle data that the authors validated. Hence the version of FDS4 used for the current study was modified so that both g RR and r LN could be input and then processed as input in all the pertinent calculations within FDS4. Both g RR and r LN can be independently defined in FDS5. Heat of Combustion The gross heat of combustion of wood burned under optimum conditions may be as high as 2.4 MJ/kg [59]. However, for the current investigation, a H C of 15 MJ/kg for wood in cribs was used based on a review of the available published data and the understanding that the

10 158 J. TRELLES AND J. R. MAWHINNEY Cumulative volume fraction F (d ) Rosin-Rammler branch.2.1 Nonsmooth log-normal branch Smooth log-normal branch.1. Measured d (µm) Figure 3. The nonsmooth (at the median drop size point) log-normal branch of cumulative volume fraction for a Marioff 4S 1MC 8MB 11 water mist nozzle better matches the data than does the smooth default of standard FDS4. (The color version of this figure is available online.) water mist would not allow the full 2.4 MJ/kg to be achieved. This value is deemed to be representative of standard severity fuel packages, mainly common wood combustibles. Oxygen Depletion It is well known that one of the suppression mechanisms of water mist is oxygen depletion. The evaporating water displaces the oxygen, resulting in lower oxygen concentration. Combustion itself consumes the available oxygen. FDS can alter the heat release rate by comparing the local oxygen concentration and temperature with an oxygen volume fraction-temperature map [56] that delimits burn and no burn zones. The heat release rate reduction algorithm is invoked at cells that bind the flame interface. As Figure 4 shows, runs performed with this oxygen depletion option turned on did not always track the input HRR because of dips related to local suppression of the HRR. (This effect was more pronounced in other tests

11 CFD Investigation of Large Scale Pallet Stack Fires Heat release rate dq/dt (MW) FDS Experimental Figure 4. In earlier fuel load modeling attempts, as is shown on the left, the oxygen depletion feature was left turned on. The effects, as are shown in the plot on the right for Test 1, were dips in the computed heat release rate when compared against the input profile. (The color version of this figure is available online.) not covered in this article.) Furthermore, the measured heat release rate reflects all the oxygen depletion that would have occurred. Hence the oxygen depletion feature was turned off for all the simulations. Computational Domain The computational domain encompassed a 14 m section of the 6 m long tunnel between stations at 32 m and 46 m. The basic numerical details are given in Table 5. The cells were uniform in each coordinate. The largest cell size, 25 mm in the x-direction, corresponds with the smallest characteristic uncertainty (x c ) in Table 4. The cell size, x, is also smaller than the 8% of x c recommended in [46]. The 1% grade along the roadway is such that the left side of Figure 5 is higher than the right side. This was approximated by tilting the acceleration of gravity, that is, g! ¼ (.981,., 9.895) m/s 2, such that g! ¼ 9.81 m/s 2. Because of the curvature of the tunnel and the need to place probe points so that conditions such as the centerline temperature variation can be displayed accurately, a FORTRAN 95 program was written in order to generate the nontrivial aspects of the tunnel geometry. Figure 5 gives an example of a generated computational domain. Most items are accurately placed to within plus or minus half the resolution given in Table 5. Nozzle and instrumentation probe locations are accurate to the precision given in the experimental final report. The line down the middle of the tunnel corresponds to the tunnel centerline. Vertical bars were used to demarcate the 1 m stations along the tunnel walls. The centerline station could be determined by drawing a line between corresponding west and east station markers. The intersection with the centerline is the station location. Regions outside of the tunnel space were

12 16 J. TRELLES AND J. R. MAWHINNEY Table 5. Summary of FDS 4..7 input parameters shared amongst the different simulations. Category Parameter Value CFD Domain Facility San Pedro de Anes Research Tunnel Simulation dimensions 14 m 23 m 5.17 m Numerical Grid dimensions cells Cell size mm Total # of cells 1,344, Gravity vector (.981,., 9.895) m/s 2 Wall boundary conditions Concrete Floor boundary conditions Concrete Ceiling boundary conditions Concrete/Promat Promatect-H Nozzle Type Marioff 4S1MD6MD(1,1RE) water mist Configuration 4 m 3.3 m grid, hybrid (1,1RE) or 1RE only Activation criteria Times as determined from experimental data Figure 5. The generic computational domain for a 14 m curved section of tunnel. The left end of the tunnel (plan south) contains the forced flow boundary condition. The right end (plan north) is open. The water mist system is situated between stations 356 m and 424 m. The diagram shows ceiling thermocouples, four thermocouple trees, and a composite fuel load centered on station 39 m. (The color version of this figure is available online.) blocked off in order to avoid calculating flows in unwanted areas. The location of items such as nozzles and thermocouple trees were based on centerline station location. Points off the centerline were measured along the radius perpendicular to the centerline. Boundary Conditions The default thermal boundary condition was based on the properties of concrete as reported in Tables 6 and 7. The section of the ceiling between stations 37 m and 42 m was protected by two layers of Promat Promatect- H [6,61] insulating board. Refer to Tables 6 and 7. The right (plan north) boundary condition is open. Gases enter and leave this area according to the

13 CFD Investigation of Large Scale Pallet Stack Fires 161 Table 6. Thermal properties of the thermally thick surfaces used in the FDS modeling. Surface name Thermal conductivity k (W/(m K)) Density q (kg/m 3 ) Specific heat c (J/(kg K)) Reference Concrete Variable 27 Variable [72] Gypsum board [73] Wood [73] PT-Steel [73] Promat Promatect-H Variable [6] Item Table 7. Thicknesses of boundary elements. Thickness (m) Tunnel concrete.4 Ceiling slabs.25 Floor.3 Promat boards.32 Gypsum board.13 Wood.2 Plate thermometer steel.7 difference between the local and the outside pressure heads. The left (plan south) boundary condition is a fixed velocity according to Table 1. Because of the curved geometry, the velocity at each cell in the y-direction would differ in its x- and y-components. Hence the left boundary was broken up into vertical cell strips as is shown in Figure 6. The entries for the velocity at the center of each cell strip were calculated and output using the aforementioned domain generation program. The specified velocity boundary condition has the drawback that it does not let back layering pass through it. However, this boundary at station 32 m was sufficiently far away from the instrumented section (between 345 m and 45 m) so that flow reversal had little impact on the results of the simulations. Because mechanical forces in the governing equations underlying FDS respond with infinite speed [52,55], the velocity boundary condition establishes itself almost immediately throughout the length of the tunnel. (FDS has a default 1 s ramp up period for numerical stability reasons.) Nonetheless there are other transients, such as the wake downstream of the pallet stack, that need time to decay to the predominant turbulent flow. For this reason, up to 1 min of delay time was provided before the fire was allowed to start burning.

14 162 J. TRELLES AND J. R. MAWHINNEY Figure 6. Diagram illustrating the method of inputting the upstream velocity boundary condition as a series of vertical strips, one cell-width wide. (The color version of this figure is available online.) Fuel Load and Pallet Supports Because of the nominal ¼ m resolution in the computational domain, the full detail of the stacked pallets could not be represented. Instead, methods were explored by which some of the effects of flow through porous media could be obtained given that FDS has no such capabilities and given the resolution used for the simulations. In the final method adopted, denoted as the top cell method, the fuel load was modeled as the main member in the pallet load with wood thermal boundary conditions (see Figures 7 and 8 for visualizations of the pallet stacks that formed the fuel array). The porosity of pallet stacks varies in the downstream direction but a uniform porosity approach was pursued. Although the sole function of the fuel load in the top cell method is to support the top cells, modeling the fuel load as a collection of main member allows air to flow through the arrangement. This addressed the concern of having an otherwise impenetrable obstacle in the center of the tunnel. Initial trials with a full height commodity found that not enough space was available for flame volume, even with the receding pallet stack. Much better results were obtained with a half height fuel load. This method allowed for more flame volume above the fire bed cells that are all located on one vertical plane. The drag across the half-height load was lower than

15 CFD Investigation of Large Scale Pallet Stack Fires 163 Figure 7. The actual pallet stacks, shown on the left, afford avenues for gas flow. (The darker pallets on the bottom are the supports. Although not clearly visible, there are layers of gypsum board separating the fuel pallets from the supporting pallets.) In order to facilitate air flow through the fuel load, the model shown on the right incorporates only the main vertical members. (The color version of this figure is available online.) that had been the case with the full-height and downstream flame lengths were longer than that was encountered experimentally. Nonetheless, this method was adopted because it produced the correct temperature range near the ceiling. To complete the fuel load obstruction model, a zero-thickness obstacle was used to model the gypsum board sheets placed on the sides and top of the pallet stand. The thermal properties for gypsum are given in Tables 6 and 7. The pallet core of this support was modeled as a collection of wooden blocks that likewise allowed air to flow through it. Heat Release Rate Heat release rates were measured in the experimental tests by instrumentation placed at the exit portal to the tunnel, which was located approximately 2 m downstream from the end of the fuel package. Hence there was an inherent time delay with all HRR measurements when compared with other instrument readings such as thermocouples. Furthermore, the HRR time histories were noisy due to normal fluctuations in instantaneous instrument readings. In order to avoid exceeding tolerances in the flame spread algorithm, the input HRR signals were smoothed. The time lag was also reconciled against other instrument readings. Refer to [4] for the details.

16 164 J. TRELLES AND J. R. MAWHINNEY Figure 8. The pallets in the support platform were modeled using a staggered block arrangement in order to differentiate them from the fuel load while allowing gases to pass through. The checker-board patterns (a consequence of the post-processor) denote the gypsum board sheets. (The color version of this figure is available online.) The flame spread methodology was as follows. The top cells on the top of the loads would start to burn according to the input HRR curve. Each cell had the same heat release rate per unit area, _Q. This quantity varied from simulation to simulation. It was determined by taking the maximum HRR achieved in a test and multiplying it by a scaling factor proportional to the cell-life-to-run-time ratio, dividing it by the number of available cells, and by the area of one top cell. In the absence of longitudinal flame spread data, each cell was given a finite life of about a quarter of the total simulation time (see Figure 9 for an example). This would create a de facto spreading front across the surface of the fuel load. The number of available top cells varied from simulation to simulation as well, being determined by the fuel array dimensions and by the percentage of the pallet load that was left standing at the termination of the test (as listed in Table 2). Once a strip of top cells ceased to burn, the supporting obstacles were removed as well. The progression was from strip center cell to cells on the left and right (in an even fashion as can be seen in Figure 1) and from the front (upstream) to the back (downstream) of the pallet load (again, refer to Figure 1).

17 CFD Investigation of Large Scale Pallet Stack Fires 165 HRRPUA dq /dt (kw/m 2 ) Ramp up time Steady burn period Ramp down time Time t (s) Figure 9. Because the input HRR was matched to a tolerance, each combustion cell could have a unique life span or reference one from an already established collection. Ramp up and ramp down periods were typically 15 s. All cells shared the same steady burn rate but this differed from one simulation to another. (The color version of this figure is available online.) Table 8 gives further data related to the top cell method. In FDS, _Q is input as the heat release rate per unit area according to the values given in Table 8. These values are, of course, much greater than the corresponding heat fluxes at the fuel surface, _q. At a surface in FDS, there is only volatilization, that is, no combustion. The actual heat comes from the flame surface that is off the body of the fuel. This flame surface is much larger than the fuel top surface area. Therefore the heat flux reaching the fire bed from the flame surface would be much less than the values given in Table 8. Internally, FDS takes _Q and the heat of combustion to determine the mass release rate per unit area at the top surface, _m given in Table 8, which is what FDS actually uses to compute the flame. It can be shown that the tabulated values of _Q are reasonable when all the heat generation is channeled through the top surface of a pallet stack. For a 2.1 m high configuration, data from Babrauskas [62] set the HRR at 6 MW. Dividing by the gross top area (i.e., not taking into account sections that did not burn out) gives a free burn heat release rate per unit area _Q fb 4 MW/m2. This will be higher in the presence of ventilation. Carvel and Beard [63] use the equation _Q v ¼ k uðuþ _Q fb to estimate the augmented HRR. The correction factor, k u, is a function of the velocity, u, and the fuel type. For u 2 m/s and a heavy goods vehicle as the fuel source, the graph in [63] gives the expectation value hk u i 3, which implies that _Q v 12 MW/m 2. Although less than the 14.8 MW/m 2 of the unsuppressed run, the 12 MW/m 2 value is comfortably within the uncertainty of the correction and is also greater than the 7.9 MW/m 2 of Test 1. Because the water mist system was active, a reduction in the HRR is to be expected.

18 166 J. TRELLES AND J. R. MAWHINNEY Figure 1. Illustration of the burning sequence of top-cells on the support structure. Cells burn from front to back along the top of the support structure only. The combustion progresses as a traveling band of open cells. Uninvolved cells are denoted with an. Cells that burn-out are also shown with an. For these cells, the supporting structures are removed in order to simulate the collapse of the fuel stacks to that extent in the downstream direction. In this case, an upstream portion of the pallet stacks never burned away: (a) s, (b) 43 s, (c) 194 s, (d) 1446 s. (The color version of this figure is available online.) An unsuppressed simulation cannot use the Test 1 HRR in a computational domain without water mist. The HRR would actually be higher because the water mist system did suppress the HRR. The Test 1 fuel load consisted of a combination of wood and plastic pallets in a configuration that was designed to produce a more severe fire than that of Test 1. Its unsuppressed HRR was estimated to have a potential maximum of 1 MW. However, with the water mist activated, the Test 1 HRR did not exceed 57.5 MW and showed no appreciable period of steady burning at this peak. Because of the absence of a period of constant HRR, the Test 1 HRR curve appears to be very close to an unsuppressed fire curve and so it

19 CFD Investigation of Large Scale Pallet Stack Fires 167 Table 8. Parameters at the fire boundary conditions ( Q _ is the heat release rate per unit area and m_ is the mass release rate per unit area). Test identifier No. of cells _Q (kw/m 2 ) _m (kg/s/m 2 ) Unsuppressed ,8.988 was chosen to be representative of an unsuppressed fire. The fuel geometry and ventilation of Test 1 were used for the unsuppressed run. The unsuppressed HRR could have been that of Test 1 until the nozzles came on, switching to the Test 1 HRR from then on. However, it was decided to use the unadulterated Test 1 HRR because it came directly from measurements. A consequence of this decision is that the two runs start to differ noticeably just before the water mist system comes on in Test 1. Watermist System Activation times for the open and thermally activated water mist nozzles were obtained from the experimental data. The sources include notes on observations made during the tests and the system water pressure plots. Even though the domain generation program creates the whole water mist system, only nozzles that activated were used in each simulation in order to minimize the calculation overhead. The pressure in the nozzle characterization files was set to match the nominal zone pressures for each test. The default policy of FDS4 is to remove droplets that have reached the floor. The opposite setting would use the droplets that reached the ground in evaporation and heat transfer calculations. This would have the desirable effect of cooler floor temperatures but at the cost of extended run times because of the increasing number of droplets that FDS4 would have to manage. Simulations were performed which maintained droplets on the floor until they completely evaporated away. The differences in the results were negligible. Hence the simulation suite used the option to remove droplets once they reached the floor (i.e., the lowest index in the z-coordinate computational grid). The spherical model for characterizing the nozzle was employed. In this methodology, droplets can be introduced through any of the user-defined solid angles that make up the sphere surrounding the nozzle (see [52] for further details). The sphere was chosen to have a radius of.2 m and was divided into 156 solid angles; 54 of the solid angles were assigned a nonzero flow value. The distribution of the droplets from the 54 solid angles was

20 168 J. TRELLES AND J. R. MAWHINNEY Table 9. Nozzle characterization parameters where m_ c is the mass flux at the center port and m_ r is the same quantity for the ring ports. Nozzle pressure (bar) T H2O (8C) u (m/s) 4S 1MD 6MD 1 _m c (kg/s/m 2 ) _m r (kg/s/m 2 ) u(m/s) 2N 1MD 6MD 1RE _m c (kg/s/m 2 ) _m r (kg/s/m 2 ) determined as follows. Both nozzles have one port on the center axis and 6 side ports (refer to Figure 1 and 2). One solid angle, situated near 458 from the south pole, was used for each side jet. These six solid angles were evenly spaced around the azimuthal angle. The 48 solid angles that cluster around the south pole of the sphere were used for the center jet. Required inputs are the initial droplet velocity and the flux through the face of each solid angle. These were determined by proportioning the flow rate of each orifice through the area perpendicular to the solid angle at a radius of.2 m from the imagined centroid of the nozzle. Table 9 details the parameters that were used to characterize the nozzles for each of the test runs. The droplet injection rate was 1 khz. The time between injections of droplets was.5 s. This means that 5 droplets were introduced from each nozzle in.5 s intervals. Spray refinement experiments were performed. It was found that reducing the time between injections of droplets improved agreement with measured data. However, substantially increasing the injection rate to 35 khz and reducing the time between injections of droplets by a factor of five (i.e., 35 droplets introduced per nozzle in.1 s intervals) increased run times from one week to one month while only nominally improving results. Instrumentation Probe points were placed at the thermocouple (TC) locations indicated in the test documentation. These included thermocouples in individual and tree arrangements, velocity probes, and plate thermometers. Refer to Figure 11 for the layout. Centerline Ceiling Thermocouples Twenty-two ceiling TCs were spaced 5 m apart along the centerline of the tunnel, between stations 345 m and 45 m. They were located.1 m below the ceiling and were labeled C1 through C22, with C1 at Station 45 and C22 at Station 345.

21 CFD Investigation of Large Scale Pallet Stack Fires 169 Center line Tcs: C1 to C22 at 5-m spacing. Appox..1-m off center-line Approx..1 m below ceiling surface TC trees at T1 and F1 upwind TC trees at F2 and T2 downwind Water mist nozzles 356-m to 426-m 42-m C7 Tree F2 435-m C4 Tree T2 45-m C1 Fuel package: 386-m to 393-m 345-m C22 Tree T1 36-m C19 Tree F1 ISOMETRIC VEIW OF INSTRUMENTATION ZONE Not to scale. Curvature not represented. Figure 11. Diagram illustrating the locations of field instrumentation in the test tunnel, excluding the exit portal gas analyzers. (The color version of this figure is available online.) Thermocouple Trees The test facility contained four thermocouple trees (TCTs). Figure 11 illustrates the locations of the T1, F1, F2, and T2 thermocouple groups. The TCTs had two points on the centerline and three to the SW and NE of the centerline. The cross-tunnel spacing was 3.2 m. The top TCs were 5.1 m above the floor and.1 m below the ceiling. The middle tier was 3.35 m above the floor and the lowest tier was 1.5 m above the floor. The T1 and T2-series contained only thermocouples. The T1-set was centered on the 345 m station. The T2-set was anchored at the m station. In the simulations, the velocity was also monitored at the T1 and T2 locations even though no bi-directional velocity probes were installed at those locations in the actual tests. The F1 and F2-locations (Figure 11) had the same arrangement of TCs as the T-series. The F1-set was centered on the 36 m station; the F2-set was anchored at the m station. Only gas temperature was monitored at these points. Plate Thermometers It was the intention of the test series to use plate thermometers (PTs) at the F1 station to determine the heat flux at various locations. Unfortunately, damage to the insulating back surfaces of the plate thermometers made the

22 17 J. TRELLES AND J. R. MAWHINNEY Figure 12. The three cubes visible behind the fuel load are the downstream plate thermometers used in the Test 1 simulation. The dark face on each block indicates the plate thermometer orientation. (The color version of this figure is available online.) measurements unreliable. Plate thermometers stabilize temperature readings during highly turbulent fire conditions. Hence they are useful for establishing performance criteria for water mist systems. It was decided to insert them in the CFD analyses. Each plate thermometer was modeled as a single-cell block. The data collection side had the properties of.7 mm thick steel with a perfectly insulated backing. The plate thermometer face had the dimensions of the corresponding computational cell surface. This is as close as the recommendations in [64] could be followed given the limitations of resolution and of the FDS model. The plate thermometer temperature is the wall temperature of the exposed face calculated by FDS, thus providing a physicsbased model for damped temperatures. In addition, at the face of each plate thermometer, the net heat flux calculated by FDS was recorded. Unlike the methodology presented in [64], where the heat fluxes were calculated from damped temperature readings, the net heat flux to a surface as calculated by FDS is undamped. Figure 12 shows the downstream plate thermometers for Test 1. BASIC VERIFICATIONS Verification and validation have been performed according to Department of Defense Guidelines [65]. In this section, certain calculated

23 CFD Investigation of Large Scale Pallet Stack Fires Speed v (m/s) Unsuppressed Test 1 Test 2 1. Test 3 Test 1 Test Time t (s) Figure 13. Simulation results for the velocities at the T1 location as functions of time. (The color version of this figure is available online.) values from the simulations of five fire tests are verified, that is, examined for reasonableness, when compared to the known or measured conditions during the fire tests. Comparison of the calculated conditions from the simulations, with corresponding test plots from the fire test instrumentation, is presented and discussed in the next major section. Tunnel Ventilation Air Velocity The ventilation air velocity in the tunnel prior to ignition of the fire was measured for each test as an average over a nine-point equal area traverse upwind of the fuel array. This average was input as a boundary condition. To verify that the simulation reflected the measured ventilation conditions, the calculated average air velocity versus time plots are shown for a particular location. The averages of the ventilation air velocities at the T1 location, 4 m upstream of the fire, are shown in Figure 13. For the lower ventilation velocities and faster fire growth rates, the back-layering could reach the T1-station. This can be seen in Figure 13 as a disruption of the otherwise steady readings. Recall that the grade of the tunnel roadway favors buoyant flow in the direction opposite to the fan velocity. The value for Run 1 in Table 1 was used for the unsuppressed run (i.e., 2 m/s). The T1 locations at which measurements were made were not the ceiling points where the back layering was strongest. The back layering reduced the cross-sectional area through which the flow associated with tunnel ventilation could travel. It is also affected by the fixed velocity boundary condition as

24 172 J. TRELLES AND J. R. MAWHINNEY (a) Heat release rate dq/dt (MW) FDS EXP (b) Heat release rate dq/dt (MW) FDS EXP Figure 14. Comparisons are provided of the input (experimental) heat release rate with that computed by FDS for Test 1 (a) and the unsuppressed run (b). The error bars associated with the uncertainty given in Table 4 are shown. (The color version of this figure is available online.) mentioned above. Hence, by continuity, the measured velocity at the stations below the back-layer increased as is evidenced by Figure 13. This implies that, even though the purpose of Figure 13 is basic verification of a velocity boundary condition, it also serves to indicate when significant back-layering arrived at station T1. Because of the high heat release rate in the unsuppressed run, the back layering reached the T1 position by 2 s and dominated the velocity readings from thereon. Conditions before 2 s show that the target velocity was achieved. In spite of the limitations of the ventilation BC, Figure 13 shows no evidence of recirculated flow reaching the T1 position. Heat Release Rates The algorithm that generated the heat release rate versus time curve for each test was designed to match the HRR curve measured during the test (smoothed and shifted as discussed above). To verify that this algorithm worked, the heat release rate calculated by FDS is compared in Figure 14 with the smoothed profile from Test 1, with the uncertainty (error bars) added, that was used to generate the input. The output is mostly within the error bars. The results for all the cases were found to be in excellent agreement. In addition, checks were performed to ensure that nozzles came on at the right time and in the right pattern for all the runs. Since all the nozzles come on at the same time for Test 1, the results are not presented here. Centerline Temperature SELECT VALIDATIONS Figure 15 shows the ceiling temperature profiles along the centerline of the tunnel at different times during Test 1, as predicted by the simulation

25 CFD Investigation of Large Scale Pallet Stack Fires 173 (a) Temperature T ( C) (b) Temperature T ( C) (c) Temperature T ( C) C22 C21 Ventilation air 2 m/s C2 C19 C18 C17 C16 C15 C14 C13 C12 C11 C1 C9 C8 C7 C Tunnel station s (m) C22 C21 Ventilation air 2 m/s C22 C21 C2 C19 C18 C17 C16 C15 C14 C13 C12 C11 C1 C2 C19 C18 C17 C16 C15 C14 C13 C12 C11 C1 C9 C8 C7 C6 C9 C8 C7 C6 C5 C5 C4 C3 C5 C4 C3 C2 C1 EXP Test 1 FDS Test 1 FDS Unsuppressed Tunnel station s (m) Ventilation air 2 m/s C4 C3 C2 C1 C2 C1 EXP Test 1 FDS Test 1 FDS Unsuppressed EXP Test1 FDS Test1 FDS Unsuppressed Tunnel station s (m) Figure 15. Three point (2 s) averages of the temperature along the tunnel centerline at the ceiling thermocouple points at the indicated times for Test 1 and the unsuppressed run are compared with the experimental results. The labels along the tops of the figures are the thermocouple designations: (a) 42 s, (b) 72 s, (c) 12 s. (The color version of this figure is available online.)

26 174 J. TRELLES AND J. R. MAWHINNEY and as measured in the fire test. The lower abscissa shows the tunnel station numbers and the upper abscissa shows the corresponding TC designation. The extent of the pallet array is indicated by the block above the lower abscissa. The larger rectangle denotes the area protected by water mist. Figure 15 gives three-point (2 s) averages of the temperature at the ceiling for three times, as predicted by the simulation for Test 1. The first time (42 s) is indicative of the highest temperature conditions just before the water mist system came on. Each successive time period was 5 min later than the previous one, and is cooler than the profile shown for 5 minutes earlier. The results clearly show that the water mist eliminated back layering upstream of station 375 m. The agreement of the back layering branches with their experimental counter points in Figure 15 is very good. The temperature at C14 was higher in the tests than in the simulation. From C13 to C1, though, the numerical results are generally higher than the test data, the dip at C11 is due to the presence of a nozzle just below the TC. Thus, prior to activation of the water mist system, there is extremely good agreement between the simulation and the test data for regions 1 or more meters from the fire location. Otherwise, the tests recorded much better temperature reduction along the ceiling than was predicted by the model. It must be noted, though, that FDS is providing dry gas temperature while the test measurements were affected by the moist environment created by the sprays and may record a wet-bulb temperature expected to be lower than the gas temperature. Another point to be noted is that the error bars shown in Figure 15 refer to the uncertainty in the simulation results only. They are not representative of the difference between test and prediction. They are an indication of how the uncertainty in the input heat release rate manifests itself in the temperature predictions. Immediately over the fire itself there are several differences between the simulation and the test results. The peak temperature measured (at C13 in the test) was above 18C whereas the peak temperature (at C12 in the simulation) was approximately C (but higher at adjacent times). The shift of peak from C13 to C12 is an artifact of the method of characterizing the fuel package, and is considered to be of minor significance. In the area directly above the fire, flame extensions impact directly on the ceiling. Temperatures of 88C or greater are deemed to represent the presence of flame [66]. The distance from the top of the fuel package to the ceiling was only 1.9 meters, while the flame height of a 2 MW fire would be expected to be in the order of 1 m. In the confinement of a tunnel, flame height converts to flame length in a horizontal direction. In the test data, three thermocouples (C12, C13, and C14) showed the direct influence of intermittent or continuous flame. As shown in Figure 16, TC C14 was about

27 CFD Investigation of Large Scale Pallet Stack Fires 175 C16 C15 Region of direct flame impingement on tunnel ceiling C14 C13 C12 Ceiling thermocouple C11 Ventilation air Figure 16. Sketch showing the relationship of ceiling thermocouples to the flame region. (The color version of this figure is available online.) 1 m upstream of the fuel package, C13 was at the mid-point of the fuel array, and C12 was approximately 1 m beyond the downstream edge of the fuel array. It is expected then that TC C13 would be the first to be touched by flame, C14 would see heat once the fire became well established, and C12 would be increasingly exposed to heat as the fire burned toward it. Comparison of the unsuppressed run with Test 1 shows universally worse conditions without water mist. The most dramatic difference is with the back layering. Without water mist, hot gases extend both upwind and downwind of the fire, making approach from both directions very hazardous, and flame spread to vehicles on either side of the fire is assured. The FDS results are conservative and trend correctly. Figure 17 shows the time series of select ceiling TCs. In general, the higher the temperature, the higher the turbulence. Note how the temperatures drop after the water mist comes on. Also notice how the droplets lead to overall noisier signals. The high frequency oscillations in these signals makes it difficult to analyze the results. Hence, piecewise Bézier-spline smoothing [67] was employed to generate the profiles in Figure 18. Overall, Figure 18(b) compares favorably with its experimental counterpart in Figure 18(a). In general, the tests saw more flame at C13 while the simulations recorded it at C12. As was mentioned above, this is a consequence of the half-height pallet load, which resulted in more flame extension down the tunnel due to overall less resistance through the pallet load. Otherwise these two curves are analogous: the water mist cannot reduce the temperature in the presence of direct flame; so the temperature is in the neighborhood of 98C. The simulated water mist system controlled the temperature better than was recorded at experimental measurement points C14, C13, and C12 and worse than was recorded at C7 and C11. In other words, the predicted temperatures at the ceiling are comparable with those measured during the test. Their arrangement differs because of the difficulty encountered in modeling the flame spread through

28 176 J. TRELLES AND J. R. MAWHINNEY (a) Temperature T ( C) C7 C11 C12 C13 C14 (b) 11 Temperature T ( C) Figure 17. Temperature histories along the tunnel centerline at the indicated instrumentation points for Test 1 and the unsuppressed run (a) FDS Test 1, (b) FDS Unsuppressed. (The color version of this figure is available online.) C7 C11 C12 C13 C14 the fuel package. Peak temperatures are best judged from Figure 17 which shows that they compare quite favorably. For the unsuppressed fire in Figure 18(c), as the fire spreads along the top of the fuel load, the temperature becomes more uniform along the ceiling centerline. In examining Figure 18, it is evident that the simulation reproduced the major cooling effects associated with the water mist acting on the fire.

29 CFD Investigation of Large Scale Pallet Stack Fires 177 (a) Temperature T ( C) (b) Temperature T ( C) (c) Temperature T ( C) Figure 18. Smoothed temperature histories along the tunnel centerline at the indicated instrumentation points for Test 1 and the unsuppressed run with error bars added are compared against the measured temperature time lines along the tunnel centerline at the indicated instrumentation points for Test 1: (a) Experimental Test 1, (b) FDS Test 1, (c) FDS Unsuppressed. (The color version of this figure is available online.) Just prior to activation of the water mist system, the temperature at TC C7, 26 m from the fire area, was measured at 358C; the simulation indicated a temperature of 38C. At the same time, TC C11 indicated temperatures just over 58C; the simulation showed approximately 558C. The differences between test and simulation temperature were within 58C. Prior to activation of the water mist, the thermocouples immediately above the fuel package in the flame zone, that is, C12 and C13, recorded temperatures C7 C11 C12 C13 C14 C7 C11 C12 C13 C14 C7 C11 C12 C13 C14

30 178 J. TRELLES AND J. R. MAWHINNEY from 88C to 18C. The simulation predicted temperatures in the same range, i.e. above 88C. Following activation of the water mist, the simulation captured the dramatic reduction in temperatures downstream of the fire, particularly for the regions not immediately in the flame zone. For TCs C1 to C7 the simulation temperatures were generally between 158C and 28C, whereas the measured values indicated temperatures between 58C and 18C. Thus, the simulation temperatures were approximately 18C higher than measured values. Given that the FDS simulation reports gas phase temperature, whereas the thermocouples record a wet-bulb temperature, it is to be expected that the simulation will predict higher temperatures than measured in the area of water spray. In the simulation, Figure 18(b) shows that only two TCs (C12 and C13) registered temperatures close to the flame temperature. It is noted that in the test (Figure 18(a)), TC C13 was apparently in direct flame through most of the test, whereas TC C14 saw more heat and C12 much less than in the simulation: the area of flame contact on the ceiling was shifted down-wind to TC C12. This shift is attributed to the approximations inherent in the topcell method. The top-cell method devised for this study succeeded in modeling the HRR versus time curve. However, details of the geometry of the flames rising through the wood pallets and the transition of flame height to flame length in the confined dimensions of the tunnel made it difficult to quantify the important parameter _Q. Notwithstanding the shift of the region of flame impact from C13 to C12, the simulation clearly illustrates that there is a region directly above the fuel array where flames will impinge on the ceiling, and where it is impractical to expect temperatures to be below the damage threshold for concrete. The model, with reasonable accuracy, defined the limits of the area of tunnel ceiling where thermal damage to structural concrete may be unavoidable. Temperature at the Thermocouple Trees Figures show the temperature at the thermocouple trees (TCTs) at T1, F1, F2, and T2. The arrangement of the figures is from the most upstream TCT (T1 in Figure 19) to the most downstream TCT (T2 in Figure 21). The trees were located near enough to ceiling stations that the ceiling TC reading was used for the ceiling center data. Because of the layering, readings at the same height are presented in the same shade with different line types. Typically, these cannot be discerned due to overlap of readings at the same level. Comparing with the experimental counterparts in Figure 19(a), the three ceiling TCs in Figure 19(b) agree very well although the numerical rate of rise is steeper. The predicted mid-height temperatures are higher than the measured values by more than 58C but the lowest rung

31 CFD Investigation of Large Scale Pallet Stack Fires 179 (a) Temperature T ( C) (b) Temperature T ( C) (c) Temperature T ( C) T1-1 C22 T1-6 T1-2 T1-4 T1-7 T1-3 T1-5 T T C22 T1-6 5 T1-2 T T1-7 T1-3 4 T1-5 T T C22 T1-6 5 T1-2 T T1-7 T1-3 4 T1-5 T Figure 19. Measured temperature histories at the T1 thermocouple tree (4 m upstream of the fire) for Test 1 are compared against the computed temperatures for Test 1 and the unsuppressed run: (a) Experimental Test 1, (b) FDS Test 1, (c) FDS Unsuppressed. (The color version of this figure is available online.)

32 18 J. TRELLES AND J. R. MAWHINNEY (a) Temperature T ( C) (b) Temperature T ( C) (c) Temperature T ( C) C19 F1-6 F1-4 F1-7 F1-3 F1-5 F F1-1 C19 F1-6 F1-2 F1-4 F1-7 F1-3 F1-5 F F1-1 C19 F1-6 F1-2 F1-4 F1-7 F1-3 F1-5 F Figure 2. Measured temperature histories at the F1 thermocouple tree (25 m upstream of the fire) for Test 1 are compared with the results of the simulations for Test 1 and the unsuppressed run: (a) Experimental Test 1, (b) FDS Test 1, (c) FDS Unsuppressed. (The color version of this figure is available online.)

33 CFD Investigation of Large Scale Pallet Stack Fires 181 (a) Temperature T ( C) (b) Temperature T ( C) (c) Temperature T ( C) F2-1 C7 F2-6 F2-2 F2-4 F2-7 F2-5 F F2-1 C7 F2-6 F2-2 F2-4 F2-7 F2-3 F2-5 F F2-1 C7 F2-6 F2-2 F2-4 F2-7 F2-3 F2-5 F Figure 21. Measured temperature histories at the F2 thermocouple tree (27 m downstream of the fire) for Test 1 are compared with the results of the simulations for Test 1 and the unsuppressed run: (a) Experimental Test 1, (b) FDS Test 1, (c) FDS Unsuppressed. (The color version of this figure is available online.)

34 182 J. TRELLES AND J. R. MAWHINNEY (a) Temperature T ( C) (b) Temperature T ( C) T2-1 C4 T2-6 T2-2 T2-4 T2-7 T2-3 T2-5 T T2-1 C4 T2-6 T2-2 T2-4 T2-7 T2-3 T2-5 T (c) Temperature T ( C) T2-1 C4 T2-6 T2-2 T2-4 T2-7 T2-3 T2-5 T Figure 22. Measured temperature histories at the T2 thermocouple tree (approximately 4 m downstream of the fire) for Test 1 (left) are compared with the results of the simulations for Test 1 and the unsuppressed run: (a) Experimental Test 1, (b) FDS Test 1, (c) FDS Unsuppressed. (The color version of this figure is available online.)

35 CFD Investigation of Large Scale Pallet Stack Fires 183 of TCs compare quite well. At F1 (Figure 2(b)), closer to the fire but still upstream, the trend is the same. At F2 (which is downstream of the fire) in Figure 21, the agreement before water mist activation is very good. After activation, the predicted ceiling TC temperatures were higher, the mid-level TC temperature agreed very well, and the simulated lowest rung TCs was better than what was measured in the tests. Overall, the comparison of the simulation and the test results far from the fire is very good. For Test 1, Figure 21(b) shows that, prior to activation of the water mist, the stratification in the tunnel is evident as is the case in the test data shown in Figure 21(a). Three ceiling elevation thermocouples recorded temperatures near 38C. Thermocouples at 1.5 and 3.15 m heights showed temperatures in the range of 5 758C. Strikingly, the simulation also showed ceiling temperatures at 38C prior to system activation, only slightly higher temperatures than was measured at the 3.15 m height. Temperatures at 1.5 m height were very close to those measured. After activation of the water mist in the fire test, temperatures at all elevations and positions at F2 were measured between 58C and 68C, indicating a fully mixed, non-stratified region over the height and width of the tunnel. In contrast, the simulation indicated a greater degree of stratification at F2 than was evident in the test data. The high temperature trace (TC C7 at the ceiling) at approximately 1758C was significantly higher than the 68C measured. However, temperatures at 1.5 and 3.15 m were within 18C of the measured values. As was discussed in relation to Figure 15, FDS predicted ceiling temperatures from 58C to 18C higher than were measured at F2. Again, the same reasons for the differences as were put forth in the previous section apply here as well. Because of the lack of back layering control in the unsuppressed fire, Figure 19(c) shows that the ceiling temperature at T1 presents a burn hazard for inhabitants at the ground. The situation only gets worse at F1 (Figure 2(c)). Downstream, Figure 21(c) shows that temperatures are as high as 98C at the F2 station which is 27 m from the center of the fuel array. At station T2, 4 m from the fire, Figure 22(c) shows moderate improvement compared to Figure 21(c) at 27 m from the fire. For the unsuppressed fire case, it is expected that even 4 m away from the fire, the tunnel lining would be exposed to temperatures high enough to cause spalling of unprotected concrete, leading to catastrophic damage to the tunnel structure. Energy Absorbed by the Mist DISCUSSION A good measure of the impact of the water mist system on the fire can be obtained by integrating the mass of water per time per area, _m H 2 O,at

36 184 J. TRELLES AND J. R. MAWHINNEY Volumetric flow rate dv/dt (L/min) At nozzles At floor Floor avg Figure 23. Volumetric flow rate of water mist reaching the floor for Test 1. The solid trace indicates the calculated volumetric flow rate of water mist reaching the floor. The dashed bar is the average over the indicated time period. The dash-dot-dot line represents the average flow rate out of all the nozzles. The difference between the dash-dot-dot and the dashed curves is due to evaporation by the fire plus losses out the open end of the tunnel. (The color version of this figure is available online.) the floor. By comparing the resulting curve with the one representing the injection at all the water mist nozzles gives an indication of how much water was evaporated by the fire. The following analysis is approximate. Droplets exiting the tunnel were not recorded. However, the energy required to elevate droplet temperatures was not taken into account either. In Figure 23, the dash-dot-dot line shows that a combined flow rate of 113 L/min was introduced into the computational domain by the 21 nozzles that activated at 7.5 min. The solid line denotes the integrated curve for the floor water flux. This tends to be a noisy signal because of the turbulence within the flow and the dispersed distribution of the droplets. Furthermore, droplets travel at different speeds through a medium that is itself in motion and encounter walls and obstructions along the way, resulting in non-uniform arrival times. The solid line is lower than would be expected by evaporation alone because some get advected out plan north opening of the tunnel. The dashed line is the average within the indicated period. The difference between the dashed and the dash-dot-dot line indicates that about 35 L/min, on average, of liquid water mist were lost to evaporation. Notice how the solid-line increases (i.e., the evaporation rate decreases) as the HRR decreases over the period from 25 min to 35 min. Using the heat of vaporization of water at atmospheric pressure (2.26 MJ/kg [68]), it can be estimated that 13 MW of

37 CFD Investigation of Large Scale Pallet Stack Fires 185 energy were absorbed, on average, due to evaporation. The water deposition rate at the floor responds inversely to the HRR, indicating that more evaporation will occur as the severity of the fire increases. This built-in response mechanism is one of the key fire protection features of water mist systems. Plate Thermometers The plate thermometers (PT) were modeled as described above. They were placed at the locations of targets (4 and 5 m downstream of the fire), and at TC tree stations 2 m upstream and downstream of the fuel array. Figure 24 shows that PT temperatures are smoother than their point-wise gas temperature counterparts. These temperatures reflect convection, radiative exchange in the presence of smoke and water mist, the cooling effects of droplets, and the response of the lumped mass. Note the steep gradient in the downstream direction. For the unsuppressed run, Figure 24(b) shows that the temperature readings at the 4 m and 5 m position essentially coincide when the heat release rate is above 5 MW. As was mentioned above, the heat flux readings (in the units of power per unit area) are direct reflections of the thermal environment. (A positive heat flux indicates a plate surface that is heating up and a negative heat flux corresponds with a surface that is cooling down. The y-axis bounds in Figure 25 are burn-pain thresholds.) At about 5 min, the heat flux measurements at the 4 and 5 m positions in Figure 25 cease their rapid growths due to the descent of the smoke layer. In addition, the PT readings for the suppressed run in Figure 25(a) are affected by the droplet stream after the water mist system comes on. (This is a circumstance not considered in [64].) Hence the appearance of negative heat fluxes soon after activation. From then on the heat fluxes are in response to the turbulent, smoke and droplet seeded flows in which the plate thermometers are situated. Figure 26(a) shows how the far PT heat fluxes rise until the HRR stabilizes and drop when the water mist comes on. The upstream PT slowly cools to no heat flux. The downstream PT reading starts to drop as the smoke concentration increases, dips negative at activation, and then registers some heat from the fire that is still burning. By the end of the run, both far PTs are registering essentially zero heat flux. At a distance of 2 m downstream of the pallet edge, the heat flux readings for the unsuppressed run in Figure 26(b) show a rapid rise until the smoke concentration becomes heavy. Figure 25(b) shows that the PT temperature is stable from about 12 min to 22 min. From then on it decreases. This is reflected in Figure 26(b) as negative heat fluxes (i.e., cooling of the lumped mass). For the upstream PT, after an

38 186 J. TRELLES AND J. R. MAWHINNEY (a) Temperature T ( C) (b) Temperature T ( C) m upstream 2 m downstream 4 m downstream 5 m downstream m upstream 2 m downstream 4 m downstream 5 m downstream Figure 24. Lumped mass plate thermometer temperatures for Test 1 and the unsuppressed run: (a) FDS Test 1, (b) FDS Unsuppressed. (The color version of this figure is available online.) initial rise, Figure 26(b) shows a gradual decay in response to the decreasing visibility associated with the unchecked back layering and to a flame front that is receding from the upstream PT location. A comparison of the Test 1 heat fluxes with their unsuppressed counterparts clearly shows how worse things would be even 2 m away from the fire without the water mist.

39 CFD Investigation of Large Scale Pallet Stack Fires 187 (a) m upstream 8 2 m downstream 1 4 m downstream 5 m downstream Net heat flux dq /dt (kw/m 2 ) Net heat flux dq /dt (kw/m 2 ) m upstream 8 2 m downstream 1 4 m downstream 5 m downstream Figure 25. Net heat fluxes at the plate thermometers: (a) Suppressed Test 1, (b) Unsuppressed Test 1. (The color version of this figure is available online.) (b) (a) 2. (b) m upstream m upstream 2. 2 m downstream 2. 2 m downstream Net heat flux dq /dt (kw/m 2 ) Figure 26. Net heat fluxes at the far plate thermometers: (a) Suppressed Test 1, (b) Unsuppressed Test 1. (The color version of this figure is available online.) Net heat flux dq /dt (kw/m 2 ) Comparison of Temperature Isosurfaces Figures 27 and 28 compare the evolution of the 18C, 358C, 58C, and 1,8C temperature envelopes for the suppressed and unsuppressed Test 1 simulations. Figure 27 shows conditions before the water mist comes on. Both simulations have back layering to the upwind boundary of the simulation at station 32 m. They also have 4 m stretches for the 358C and 1 m stretches of the 58C temperature envelopes. (These differ because the HRR growths differ.) At 8 min (3 s after the water mist activated), the water mist had completely removed the back layering. At 1 min, the suppressed fire was near its peak. Figure 28 shows that the 18C envelope is no longer in contact with the open boundary of the CFD model. By contrast, the unsuppressed simulation has a sizable pocket with temperatures at or above 18C. (Note that the bottom image is showing a combination of the isosurfaces and the water mist droplets.)

40 188 J. TRELLES AND J. R. MAWHINNEY (a) :7:3. (b) :7:3. Figure 27. Comparison of temperature isosurfaces for the two tunnel simulations at 7.5 min (before activation of water mist in Test 1). The values, from outermost envelope to innermost, are 18C, 358C, 58C, and 18C: (a) Unsuppressed Test 1, (b) Suppressed Test 1. (The color version of this figure is available online.) (a) :1:. (b) :1:. Figure 28. Comparison of temperature isosurfaces at 1 min. The values, from outermost envelope to innermost, are 18C, 358C, 58C, and 18C: (a) Unsuppressed Test 1, (b) Suppressed Test 1. (The color version of this figure is available online.) Floor Surface Temperature Contours The surface floor temperature at 253 s is shown in Figure 29. This is not the temperature of the gas but the temperature response of the concrete in the floor to the flames above. Figure 29 shows clear evidence of shadowing in the temperature footprint. The temperature is cool nearest to the pallet stacks. It then increases along the tunnel centerline, peaks, and then drops again. Figure 29 also shows that the highest temperatures are located off the tunnel centerline because of the turbulence of the downstream flames.

41 CFD Investigation of Large Scale Pallet Stack Fires 189 Bndry temp C Figure 29. These contours of the floor surface temperature (at 253 s for the unsuppressed 57.5 MW fire) show the phenomenon of shadowing in the lee of a burning pallet stack whereby the region of greatest exposure is some distance away from the end of the fuel array. (The color version of this figure is available online.) CONCLUSIONS The goal of this investigation was to increase the value of and confidence in the findings of an experimental series [2] concerned with pallet stack fires (representing very large heavy goods vehicle fires) in ventilated tunnels protected by water mist systems. CFD simulations were employed as the primary means of inquiry. The CFD study required the development of a flame spread algorithm of sufficient flexibility to reasonably reproduce experimental temperature measurements. The resulting top cell method did so while providing other necessary characteristics such as sufficient flame volume, flow avenues through the fuel load, and sequential collapse of the modeled pallet stacks. The CFD model was validated against data collected during the San Pedro test series. This FDS 4..7 model successfully predicted conditions such as the following: 1. Ceiling temperatures along the centerline of the tunnel prior to water mist activation to within 58C. 2. Back layering phenomena at different ventilation air velocities before and after activation of the water mist.